nsz Git - libfirm/blob - ir/be/becopyilp2.c

   1 /*
   2  * Copyright (C) 1995-2008 University of Karlsruhe.  All right reserved.
   3  *
   4  * This file is part of libFirm.
   5  *
   6  * This file may be distributed and/or modified under the terms of the
   7  * GNU General Public License version 2 as published by the Free Software
   8  * Foundation and appearing in the file LICENSE.GPL included in the
   9  * packaging of this file.
  10  *
  11  * Licensees holding valid libFirm Professional Edition licenses may use
  12  * this file in accordance with the libFirm Commercial License.
  13  * Agreement provided with the Software.
  14  *
  15  * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
  16  * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
  17  * PURPOSE.
  18  */
  19
  20 /**
  21  * @file
  22  * @brief       ILP based copy minimization.
  23  * @author      Daniel Grund
  24  * @date        28.02.2006
  25  *
  26  * ILP formalization using G=(V, E, Q):
  27  *  - 2 class of variables: coloring vars x_ic   and   equal color vars y_ij
  28  *  - Path constraints
  29  *  - Clique-star constraints
  30  *
  31  *
  32  * \min \sum_{ (i,j) \in Q }  w_ij y_ij
  33  *
  34  *     \sum_c x_nc           =  1           n \in N, c \in C
  35  *
  36  *     x_nc                  =  0           n \in N, c \not\in C(n)
  37  *
  38  *     \sum x_nc            <=  1           x_nc \in Clique \in AllCliques,  c \in C
  39  *
  40  *     \sum_{e \in p} y_e   >=  1           p \in P      path constraints
  41  *
  42  *     \sum_{e \in cs} y_e  >= |cs| - 1     cs \in CP    clique-star constraints
  43  *
  44  *     x_nc, y_ij \in N,   w_ij \in R^+
  45  */
  46 #include "config.h"
  47
  48 #include "bitset.h"
  49 #include "raw_bitset.h"
  50 #include "pdeq.h"
  51
  52 #include "util.h"
  53 #include "irgwalk.h"
  54 #include "becopyilp_t.h"
  55 #include "beifg.h"
  56 #include "besched.h"
  57 #include "bemodule.h"
  58
  59 #define DEBUG_LVL 1
  60
  61 typedef struct local_env_t {
  62         int first_x_var, last_x_var;
  63         int n_colors;
  64         bitset_t *normal_colors;
  65         pmap *nr_2_irn;
  66         DEBUG_ONLY(firm_dbg_module_t *dbg;)
  67 } local_env_t;
  68
  69 static void build_coloring_cstr(ilp_env_t *ienv)
  70 {
  71         be_ifg_t *ifg     = ienv->co->cenv->ifg;
  72         nodes_iter_t iter;
  73         bitset_t *colors;
  74         ir_node *irn;
  75         char buf[16];
  76
  77         colors = bitset_alloca(arch_register_class_n_regs(ienv->co->cls));
  78
  79         be_ifg_foreach_node(ifg, &iter, irn)
  80                 if (!sr_is_removed(ienv->sr, irn)) {
  81                         size_t col;
  82                         int cst_idx;
  83                         const arch_register_req_t *req;
  84                         int curr_node_color = get_irn_col(irn);
  85                         int node_nr = (int)get_irn_idx(irn);
  86                         local_env_t *lenv = ienv->env;
  87
  88                         pmap_insert(lenv->nr_2_irn, INT_TO_PTR(node_nr), irn);
  89
  90                         req = arch_get_irn_register_req(irn);
  91
  92                         bitset_clear_all(colors);
  93
  94                         /* get assignable colors */
  95                         if (arch_register_req_is(req, limited)) {
  96                                 rbitset_copy_to_bitset(req->limited, colors);
  97                         } else {
  98                                 bitset_copy(colors, lenv->normal_colors);
  99                         }
 100
 101                         /* add the coloring constraint */
 102                         cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_equal, 1.0);
 103
 104                         bitset_foreach(colors, col) {
 105                                 int var_idx = lpp_add_var(ienv->lp, name_cdd(buf, 'x', node_nr, (int)col), lpp_binary, 0.0);
 106                                 lpp_set_start_value(ienv->lp, var_idx, (col == (unsigned) curr_node_color) ? 1.0 : 0.0);
 107                                 lpp_set_factor_fast(ienv->lp, cst_idx, var_idx, 1);
 108
 109                                 lenv->last_x_var = var_idx;
 110                                 if (lenv->first_x_var == -1)
 111                                         lenv->first_x_var = var_idx;
 112                         }
 113
 114                         /* add register constraint constraints */
 115                         bitset_foreach_clear(colors, col) {
 116                                 int cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_equal, 0.0);
 117                                 int var_idx = lpp_add_var(ienv->lp, name_cdd(buf, 'x', node_nr, (int)col), lpp_binary, 0.0);
 118                                 lpp_set_start_value(ienv->lp, var_idx, 0.0);
 119                                 lpp_set_factor_fast(ienv->lp, cst_idx, var_idx, 1);
 120
 121                                 lenv->last_x_var = var_idx;
 122                         }
 123                 }
 124 }
 125
 126 static void build_interference_cstr(ilp_env_t *ienv)
 127 {
 128         lpp_t       *lpp      = ienv->lp;
 129         local_env_t *lenv     = ienv->env;
 130         be_ifg_t    *ifg      = ienv->co->cenv->ifg;
 131         int          n_colors = lenv->n_colors;
 132         cliques_iter_t iter;
 133         ir_node    **clique   = ALLOCAN(ir_node*, n_colors);
 134         int          size;
 135         int          col;
 136         int          i;
 137
 138         char buf[16];
 139
 140         /* for each maximal clique */
 141         be_ifg_foreach_clique(ifg, &iter, clique, &size) {
 142                 int realsize = 0;
 143
 144                 for (i=0; i<size; ++i)
 145                         if (!sr_is_removed(ienv->sr, clique[i]))
 146                                 ++realsize;
 147
 148                 if (realsize < 2)
 149                         continue;
 150
 151                 /* for all colors */
 152                 for (col=0; col<n_colors; ++col) {
 153                         int cst_idx = lpp_add_cst(lpp, NULL, lpp_less_equal, 1.0);
 154
 155                         /* for each member of this clique */
 156                         for (i=0; i<size; ++i) {
 157                                 ir_node *irn = clique[i];
 158
 159                                 if (!sr_is_removed(ienv->sr, irn)) {
 160                                         int var_idx = lpp_get_var_idx(lpp, name_cdd(buf, 'x', (int)get_irn_idx(irn), col));
 161                                         lpp_set_factor_fast(lpp, cst_idx, var_idx, 1);
 162                                 }
 163                         }
 164                 }
 165         }
 166 }
 167
 168
 169 /**
 170  * TODO: Remove the dependency of the opt-units data structure
 171  *       by walking over all affinity edges. Graph structure
 172  *       does not provide this walker, yet.
 173  */
 174 static void build_affinity_cstr(ilp_env_t *ienv)
 175 {
 176         local_env_t *lenv = ienv->env;
 177         int n_colors      = lenv->n_colors;
 178         unit_t *curr;
 179
 180         /* for all optimization units */
 181         list_for_each_entry(unit_t, curr, &ienv->co->units, units) {
 182                 ir_node *root, *arg;
 183                 int root_nr, arg_nr, i, col, y_idx, root_idx, arg_idx;
 184                 char buf[16];
 185                 int root_col, arg_col;
 186
 187                 root = curr->nodes[0];
 188                 root_nr = (int) get_irn_idx(root);
 189                 root_col = get_irn_col(root);
 190
 191                 for (i = 1; i < curr->node_count; ++i) {
 192                         arg = curr->nodes[i];
 193                         arg_nr = (int) get_irn_idx(arg);
 194                         arg_col = get_irn_col(arg);
 195
 196                         /* add a new affinity variable */
 197                         y_idx = lpp_add_var(ienv->lp, name_cdd_sorted(buf, 'y', root_nr, arg_nr), lpp_binary, curr->costs[i]);
 198                         lpp_set_start_value(ienv->lp, y_idx, (root_col==arg_col) ? 0.0 : 1.0);
 199
 200                         /* add constraints relating the affinity var to the color vars */
 201                         for (col=0; col<n_colors; ++col) {
 202                                 int cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_less_equal, 0.0);
 203                                 root_idx = lpp_get_var_idx(ienv->lp, name_cdd(buf, 'x', root_nr, col));
 204                                 arg_idx  = lpp_get_var_idx(ienv->lp, name_cdd(buf, 'x', arg_nr,  col));
 205
 206                                 lpp_set_factor_fast(ienv->lp, cst_idx, root_idx,  1.0);
 207                                 lpp_set_factor_fast(ienv->lp, cst_idx, arg_idx,  -1.0);
 208                                 lpp_set_factor_fast(ienv->lp, cst_idx, y_idx, -1.0);
 209                         }
 210                 }
 211         }
 212 }
 213
 214 /**
 215  * Helping stuff for build_clique_star_cstr
 216  */
 217 typedef struct edge_t {
 218         ir_node *n1, *n2;
 219 } edge_t;
 220
 221 static int compare_edge_t(const void *k1, const void *k2, size_t size)
 222 {
 223         const edge_t *e1 = k1;
 224         const edge_t *e2 = k2;
 225         (void) size;
 226
 227         return ! (e1->n1 == e2->n1   &&   e1->n2 == e2->n2);
 228 }
 229
 230 #define HASH_EDGE(e) (hash_irn((e)->n1) ^ hash_irn((e)->n2))
 231
 232 static inline edge_t *add_edge(set *edges, ir_node *n1, ir_node *n2, int *counter)
 233 {
 234         edge_t new_edge;
 235
 236         if (PTR_TO_INT(n1) < PTR_TO_INT(n2)) {
 237                 new_edge.n1 = n1;
 238                 new_edge.n2 = n2;
 239         } else {
 240                 new_edge.n1 = n2;
 241                 new_edge.n2 = n1;
 242         }
 243         (*counter)++;
 244         return set_insert(edges, &new_edge, sizeof(new_edge), HASH_EDGE(&new_edge));
 245 }
 246
 247 static inline edge_t *find_edge(set *edges, ir_node *n1, ir_node *n2)
 248 {
 249         edge_t new_edge;
 250
 251         if (PTR_TO_INT(n1) < PTR_TO_INT(n2)) {
 252                 new_edge.n1 = n1;
 253                 new_edge.n2 = n2;
 254         } else {
 255                 new_edge.n1 = n2;
 256                 new_edge.n2 = n1;
 257         }
 258         return set_find(edges, &new_edge, sizeof(new_edge), HASH_EDGE(&new_edge));
 259 }
 260
 261 static inline void remove_edge(set *edges, ir_node *n1, ir_node *n2, int *counter)
 262 {
 263         edge_t new_edge, *e;
 264
 265         if (PTR_TO_INT(n1) < PTR_TO_INT(n2)) {
 266                 new_edge.n1 = n1;
 267                 new_edge.n2 = n2;
 268         } else {
 269                 new_edge.n1 = n2;
 270                 new_edge.n2 = n1;
 271         }
 272         e = set_find(edges, &new_edge, sizeof(new_edge), HASH_EDGE(&new_edge));
 273         if (e) {
 274                 e->n1 = NULL;
 275                 e->n2 = NULL;
 276                 (*counter)--;
 277         }
 278 }
 279
 280 #define pset_foreach(pset, irn)  for (irn=pset_first(pset); irn; irn=pset_next(pset))
 281
 282 /**
 283  * Search for an interference clique and an external node
 284  * with affinity edges to all nodes of the clique.
 285  * At most 1 node of the clique can be colored equally with the external node.
 286  */
 287 static void build_clique_star_cstr(ilp_env_t *ienv)
 288 {
 289         affinity_node_t *aff;
 290
 291         /* for each node with affinity edges */
 292         co_gs_foreach_aff_node(ienv->co, aff) {
 293                 struct obstack ob;
 294                 neighb_t *nbr;
 295                 const ir_node *center = aff->irn;
 296                 ir_node **nodes;
 297                 set *edges;
 298                 int i, o, n_nodes, n_edges;
 299
 300                 if (arch_irn_is_ignore(aff->irn))
 301                         continue;
 302
 303                 obstack_init(&ob);
 304                 edges = new_set(compare_edge_t, 8);
 305
 306                 /* get all affinity neighbours */
 307                 n_nodes = 0;
 308                 co_gs_foreach_neighb(aff, nbr) {
 309                         if (!arch_irn_is_ignore(nbr->irn)) {
 310                                 obstack_ptr_grow(&ob, nbr->irn);
 311                                 ++n_nodes;
 312                         }
 313                 }
 314                 nodes = obstack_finish(&ob);
 315
 316                 /* get all interference edges between these */
 317                 n_edges = 0;
 318                 for (i=0; i<n_nodes; ++i)
 319                         for (o=0; o<i; ++o)
 320                                 if (be_ifg_connected(ienv->co->cenv->ifg, nodes[i], nodes[o]))
 321                                         add_edge(edges, nodes[i], nodes[o], &n_edges);
 322
 323                 /* cover all these interference edges with maximal cliques */
 324                 while (n_edges) {
 325                         edge_t *e;
 326                         pset *clique = pset_new_ptr(8);
 327                         int growed;
 328
 329                         /* get 2 starting nodes to form a clique */
 330                         for (e=set_first(edges); !e->n1; e=set_next(edges)) {
 331                         }
 332
 333                         /* we could be stepped out of the loop before the set iterated to the end */
 334                         set_break(edges);
 335
 336                         pset_insert_ptr(clique, e->n1);
 337                         pset_insert_ptr(clique, e->n2);
 338                         remove_edge(edges, e->n1, e->n2, &n_edges);
 339
 340                         /* while the clique is growing */
 341                         do {
 342                                 growed = 0;
 343
 344                                 /* search for a candidate to extend the clique */
 345                                 for (i=0; i<n_nodes; ++i) {
 346                                         ir_node *member, *cand = nodes[i];
 347                                         int is_cand;
 348
 349                                         /* if its already in the clique try the next */
 350                                         if (pset_find_ptr(clique, cand))
 351                                                 continue;
 352
 353                                         /* are there all necessary interferences? */
 354                                         is_cand = 1;
 355                                         pset_foreach(clique, member) {
 356                                                 if (!find_edge(edges, cand, member)) {
 357                                                         is_cand = 0;
 358                                                         pset_break(clique);
 359                                                         break;
 360                                                 }
 361                                         }
 362
 363                                         /* now we know if we have a clique extender */
 364                                         if (is_cand) {
 365                                                 /* first remove all covered edges */
 366                                                 pset_foreach(clique, member)
 367                                                         remove_edge(edges, cand, member, &n_edges);
 368
 369                                                 /* insert into clique */
 370                                                 pset_insert_ptr(clique, cand);
 371                                                 growed = 1;
 372                                                 break;
 373                                         }
 374                                 }
 375                         } while (growed);
 376
 377                         /* now the clique is maximal. Finally add the constraint */
 378                         {
 379                                 ir_node *member;
 380                                 int var_idx, cst_idx, center_nr, member_nr;
 381                                 char buf[16];
 382
 383                                 cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_greater_equal, pset_count(clique)-1);
 384                                 center_nr = get_irn_idx(center);
 385
 386                                 pset_foreach(clique, member) {
 387                                         member_nr = get_irn_idx(member);
 388                                         var_idx = lpp_get_var_idx(ienv->lp, name_cdd_sorted(buf, 'y', center_nr, member_nr));
 389                                         lpp_set_factor_fast(ienv->lp, cst_idx, var_idx, 1.0);
 390                                 }
 391                         }
 392
 393                         del_pset(clique);
 394                 }
 395
 396                 del_set(edges);
 397                 obstack_free(&ob, NULL);
 398         }
 399 }
 400
 401
 402 static void extend_path(ilp_env_t *ienv, pdeq *path, const ir_node *irn)
 403 {
 404         be_ifg_t *ifg = ienv->co->cenv->ifg;
 405         int i, len;
 406         ir_node **curr_path;
 407         affinity_node_t *aff;
 408         neighb_t *nbr;
 409
 410         /* do not walk backwards or in circles */
 411         if (pdeq_contains(path, irn))
 412                 return;
 413
 414         if (arch_irn_is_ignore(irn))
 415                 return;
 416
 417         /* insert the new irn */
 418         pdeq_putr(path, irn);
 419
 420
 421
 422         /* check for forbidden interferences */
 423         len       = pdeq_len(path);
 424         curr_path = ALLOCAN(ir_node*, len);
 425         pdeq_copyl(path, (const void **)curr_path);
 426
 427         for (i=1; i<len; ++i)
 428                 if (be_ifg_connected(ifg, irn, curr_path[i]))
 429                         goto end;
 430
 431
 432
 433         /* check for terminating interference */
 434         if (be_ifg_connected(ifg, irn, curr_path[0])) {
 435
 436                 /* One node is not a path. */
 437                 /* And a path of length 2 is covered by a clique star constraint. */
 438                 if (len > 2) {
 439                         /* finally build the constraint */
 440                         int cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_greater_equal, 1.0);
 441                         for (i=1; i<len; ++i) {
 442                                 char buf[16];
 443                                 int nr_1    = get_irn_idx(curr_path[i-1]);
 444                                 int nr_2    = get_irn_idx(curr_path[i]);
 445                                 int var_idx = lpp_get_var_idx(ienv->lp, name_cdd_sorted(buf, 'y', nr_1, nr_2));
 446                                 lpp_set_factor_fast(ienv->lp, cst_idx, var_idx, 1.0);
 447                         }
 448                 }
 449
 450                 /* this path cannot be extended anymore */
 451                 goto end;
 452         }
 453
 454
 455
 456         /* recursively extend the path */
 457         aff = get_affinity_info(ienv->co, irn);
 458         co_gs_foreach_neighb(aff, nbr)
 459                 extend_path(ienv, path, nbr->irn);
 460
 461
 462 end:
 463         /* remove the irn */
 464         pdeq_getr(path);
 465
 466 }
 467
 468 /**
 469  *  Search a path of affinity edges, whose ends are connected
 470  *  by an interference edge and there are no other interference
 471  *  edges in between.
 472  *  Then at least one of these affinity edges must break.
 473  */
 474 static void build_path_cstr(ilp_env_t *ienv)
 475 {
 476         affinity_node_t *aff_info;
 477
 478         /* for each node with affinity edges */
 479         co_gs_foreach_aff_node(ienv->co, aff_info) {
 480                 pdeq *path = new_pdeq();
 481
 482                 extend_path(ienv, path, aff_info->irn);
 483
 484                 del_pdeq(path);
 485         }
 486 }
 487
 488 static void ilp2_build(ilp_env_t *ienv)
 489 {
 490         int lower_bound;
 491
 492         ienv->lp = lpp_new(ienv->co->name, lpp_minimize);
 493         build_coloring_cstr(ienv);
 494         build_interference_cstr(ienv);
 495         build_affinity_cstr(ienv);
 496         build_clique_star_cstr(ienv);
 497         build_path_cstr(ienv);
 498
 499         lower_bound = co_get_lower_bound(ienv->co) - co_get_inevit_copy_costs(ienv->co);
 500         lpp_set_bound(ienv->lp, lower_bound);
 501 }
 502
 503 static void ilp2_apply(ilp_env_t *ienv)
 504 {
 505         local_env_t *lenv = ienv->env;
 506         int i;
 507
 508         /* first check if there was sth. to optimize */
 509         if (lenv->first_x_var >= 0) {
 510                 int              count = lenv->last_x_var - lenv->first_x_var + 1;
 511                 double          *sol   = XMALLOCN(double, count);
 512                 lpp_sol_state_t  state = lpp_get_solution(ienv->lp, sol, lenv->first_x_var, lenv->last_x_var);
 513
 514                 if (state != lpp_optimal) {
 515                         printf("WARNING %s: Solution state is not 'optimal': %d\n",
 516                                ienv->co->name, (int)state);
 517                         if (state < lpp_feasible) {
 518                                 panic("Copy coalescing solution not feasible!");
 519                         }
 520                 }
 521
 522                 for (i=0; i<count; ++i) {
 523                         int nodenr, color;
 524                         char var_name[16];
 525
 526                         if (sol[i] > 1-EPSILON) { /* split variable name into components */
 527                                 lpp_get_var_name(ienv->lp, lenv->first_x_var+i, var_name, sizeof(var_name));
 528
 529                                 if (sscanf(var_name, "x_%d_%d", &nodenr, &color) == 2) {
 530                                         ir_node *irn = pmap_get(lenv->nr_2_irn, INT_TO_PTR(nodenr));
 531                                         assert(irn && "This node number must be present in the map");
 532
 533                                         set_irn_col(ienv->co, irn, color);
 534                                 } else
 535                                         assert(0 && "This should be a x-var");
 536                         }
 537                 }
 538
 539                 xfree(sol);
 540         }
 541
 542 #ifdef COPYOPT_STAT
 543         /* TODO adapt to multiple possible ILPs */
 544         copystat_add_ilp_time((int)(1000.0*lpp_get_sol_time(pi->curr_lp)));  //now we have ms
 545         copystat_add_ilp_vars(lpp_get_var_count(pi->curr_lp));
 546         copystat_add_ilp_csts(lpp_get_cst_count(pi->curr_lp));
 547         copystat_add_ilp_iter(lpp_get_iter_cnt(pi->curr_lp));
 548 #endif
 549 }
 550
 551 /**
 552  * Solves the problem using mixed integer programming
 553  * @returns 1 iff solution state was optimal
 554  * Uses the OU and the GRAPH data structure
 555  * Dependency of the OU structure can be removed
 556  */
 557 static int co_solve_ilp2(copy_opt_t *co)
 558 {
 559         lpp_sol_state_t sol_state;
 560         ilp_env_t *ienv;
 561         local_env_t my;
 562
 563         ASSERT_OU_AVAIL(co); //See build_clique_st
 564         ASSERT_GS_AVAIL(co);
 565
 566         my.first_x_var = -1;
 567         my.last_x_var  = -1;
 568         my.nr_2_irn    = pmap_create();
 569         FIRM_DBG_REGISTER(my.dbg, "firm.be.coilp2");
 570
 571         my.normal_colors = bitset_alloca(arch_register_class_n_regs(co->cls));
 572         bitset_clear_all(my.normal_colors);
 573         be_put_allocatable_regs(co->irg, co->cls, my.normal_colors);
 574         my.n_colors = bitset_popcount(my.normal_colors);
 575
 576         ienv = new_ilp_env(co, ilp2_build, ilp2_apply, &my);
 577
 578         sol_state = ilp_go(ienv);
 579
 580         pmap_destroy(my.nr_2_irn);
 581         free_ilp_env(ienv);
 582
 583         return sol_state == lpp_optimal;
 584 }
 585
 586 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_copyilp2)
 587 void be_init_copyilp2(void)
 588 {
 589         static co_algo_info copyheur = {
 590                 co_solve_ilp2, 1
 591         };
 592
 593         be_register_copyopt("ilp", &copyheur);
 594 }